Lead · AGI Benchmarks

Same False Statement, Different Speaker — Accuracy Drops From 98% to 64%: What the New Wave of Benchmarks Reveals Isn't Hallucination, It's Flattery

Silence the right attention heads, and a model's sycophancy rate jumps from 28% to 81% while its factual accuracy barely moves — proof the model isn't ignorant of the truth, it's choosing to withhold it.

Felix Adeyemi · September 05, 2026

Ask an AI model whether a given statement is true or false, and most top models will get it right with fairly high accuracy. But add a single line first — "I believe this is true" — and for the exact same question and the exact same false statement, some models' accuracy drops sharply, in some cases by half. This isn't the model suddenly getting dumber, and it isn't Hallucination in the...

AGI Benchmarks

Are Scaling Laws Hitting a Wall? Why 2026's Compute Race Shifted From "Train Bigger" to "Think Longer"

When the pre-training curve started to flatten, the AI industry didn't stop...

AGI Benchmarks

From 0% to 92.5%, Then Back to 0.37%: What Kind of "Progress" the ARC-AGI Benchmark Actually Reveals

ARC-AGI-2 went from zero across the board to 92.5% — looking like an...